CS 267: Applications of Parallel Computers. Load Balancing

Transcription

1 CS 267: Applications of Parallel Computers Load Balancing Kathy Yelick 04/18/2007 CS267 Lecture 24 1

2 Outline Motivation for Load Balancing Recall graph partitioning as load balancing technique Overview of load balancing problems, as determined by Task costs Task dependencies Locality needs Spectrum of solutions Static - all information available before starting Semi-Static - some info before starting Dynamic - little or no info before starting Survey of solutions How each one works Theoretical bounds, if any When to use it 04/18/2007 CS267 Lecture 24 2

3 Load Imbalance in Parallel Applications The primary sources of inefficiency in parallel codes: Poor single processor performance Typically in the memory system Too much parallelism overhead Thread creation, synchronization, communication Load imbalance Different amounts of work across processors Computation and communication Different speeds (or available resources) for the processors Possibly due to load on the machine How to recognizing load imbalance Time spent at synchronization is high and is uneven across processors, but not always so simple 04/18/2007 CS267 Lecture 24 3

4 Measuring Load Imbalance Challenges: Can be hard to separate from high synch overhead Especially subtle if not bulk-synchronous Spin locks can make synchronization look like useful work Note that imbalance may change over phases Insufficient parallelism always leads to load imbalance Tools like TAU can help (acts.nersc.gov) 04/18/2007 CS267 Lecture 24 4

5 Review of Graph Partitioning Partition G(N,E) so that N = N 1 U U N p, with each N i ~ N /p As few edges connecting different N i and N k as possible If N = {tasks}, each unit cost, edge e=(i,j) means task i has to communicate with task j, then partitioning means balancing the load, i.e. each N i ~ N /p minimizing communication volume Optimal graph partitioning is NP complete, so we use heuristics (see earlier lectures) Spectral Kernighan-Lin Multilevel Speed of partitioner trades off with quality of partition Better load balance costs more; may or may not be worth it Need to know tasks, communication pattern before starting What if you don t? 04/18/2007 CS267 Lecture 24 5

6 Load Balancing Overview Load balancing differs with properties of the tasks (chunks of work): Tasks costs Do all tasks have equal costs? If not, when are the costs known? Before starting, when task created, or only when task ends Task dependencies Can all tasks be run in any order (including parallel)? If not, when are the dependencies known? Locality Before starting, when task created, or only when task ends Is it important for some tasks to be scheduled on the same processor (or nearby) to reduce communication cost? When is the information about communication known? 04/18/2007 CS267 Lecture 24 6

7 Task Cost Spectrum, search 04/18/2007 CS267 Lecture 24 7

8 Task Dependency Spectrum 04/18/2007 CS267 Lecture 24 8

9 Task Locality Spectrum (Communication) 04/18/2007 CS267 Lecture 24 9

10 Spectrum of Solutions A key question is when certain information about the load balancing problem is known. Leads to a spectrum of solutions: Static scheduling. All information is available to scheduling algorithm, which runs before any real computation starts. Off-line algorithms, eg graph partitioning Semi-static scheduling. Information may be known at program startup, or the beginning of each timestep, or at other welldefined points. Offline algorithms may be used even though the problem is dynamic. eg Kernighan-Lin Dynamic scheduling. Information is not known until midexecution. On-line algorithms 04/18/2007 CS267 Lecture 24 10

11 Dynamic Load Balancing Motivation for dynamic load balancing Search algorithms as driving example Centralized load balancing Overview Special case for schedule independent loop iterations Distributed load balancing Overview Engineering Theoretical results Example scheduling problem: mixed parallelism Demonstrate use of coarse performance models 04/18/2007 CS267 Lecture 24 11

12 Search Search problems are often: Computationally expensive Have very different parallelization strategies than physical simulations. Require dynamic load balancing Examples: Optimal layout of VLSI chips Robot motion planning Chess and other games (N-queens) Speech processing Constructing phylogeny tree from set of genes 04/18/2007 CS267 Lecture 24 12

13 Example Problem: Tree Search In Tree Search the tree unfolds dynamically May be a graph if there are common sub-problems along different paths Graphs unlike meshes which are precomputed and have no ordering constraints Terminal node (non-goal) Non-terminal node Terminal node (goal) 04/18/2007 CS267 Lecture 24 13

14 Sequential Search Algorithms Depth-first search (DFS) Simple backtracking Search to bottom, backing up to last choice if necessary Depth-first branch-and-bound Keep track of best solution so far ( bound ) Cut off sub-trees that are guaranteed to be worse than bound Iterative Deepening Choose a bound on search depth, d and use DFS up to depth d If no solution is found, increase d and start again Iterative deepening A* uses a lower bound estimate of cost-tosolution as the bound Breadth-first search (BFS) Search across a given level in the tree 04/18/2007 CS267 Lecture 24 14

15 Depth vs Breadth First Search DFS with Explicit Stack Put root into Stack Stack is data structure where items added to and removed from the top only While Stack not empty If node on top of Stack satisfies goal of search, return result, else Mark node on top of Stack as searched If top of Stack has an unsearched child, put child on top of Stack, else remove top of Stack BFS with Explicit Queue Put root into Queue Queue is data structure where items added to end, removed from front While Queue not empty If node at front of Queue satisfies goal of search, return result, else Mark node at front of Queue as searched If node at front of Queue has any unsearched children, put them all at end of Queue Remove node at front from Queue 04/18/2007 CS267 Lecture 24 15

16 Parallel Search Consider simple backtracking search Try static load balancing: spawn each new task on an idle processor, until all have a subtree Load balance on 2 processors Load balance on 4 processors We can and should do better than this 04/18/2007 CS267 Lecture 24 16

17 Centralized Scheduling Keep a queue of task waiting to be done May be done by manager task Or a shared data structure protected by locks worker worker worker Task Queue worker worker worker 04/18/2007 CS267 Lecture 24 17

18 Centralized Task Queue: Scheduling Loops When applied to loops, often called self scheduling: Tasks may be range of loop indices to compute Assumes independent iterations Loop body has unpredictable time (branches) or the problem is not interesting Originally designed for: Scheduling loops by compiler (or runtime-system) Original paper by Tang and Yew, ICPP 1986 This is: Dynamic, online scheduling algorithm Good for a small number of processors (centralized) Special case of task graph independent tasks, known at once 04/18/2007 CS267 Lecture 24 18

19 Variations on Self-Scheduling Typically, don t want to grab smallest unit of parallel work, e.g., a single iteration Too much contention at shared queue Instead, choose a chunk of tasks of size K. If K is large, access overhead for task queue is small If K is small, we are likely to have even finish times (load balance) (at least) Four Variations: 1. Use a fixed chunk size 2. Guided self-scheduling 3. Tapering 4. Weighted Factoring 04/18/2007 CS267 Lecture 24 19

20 Variation 1: Fixed Chunk Size Kruskal and Weiss give a technique for computing the optimal chunk size Requires a lot of information about the problem characteristics e.g., task costs as well as number Not very useful in practice. Task costs must be known at loop startup time E.g., in compiler, all branches be predicted based on loop indices and used for task cost estimates 04/18/2007 CS267 Lecture 24 20

21 Variation 2: Guided Self-Scheduling Idea: use larger chunks at the beginning to avoid excessive overhead and smaller chunks near the end to even out the finish times. The chunk size K i at the i th access to the task pool is given by ceiling(r i /p) where R i is the total number of tasks remaining and p is the number of processors See Polychronopolous, Guided Self-Scheduling: A Practical Scheduling Scheme for Parallel Supercomputers, IEEE Transactions on Computers, Dec /18/2007 CS267 Lecture 24 21

22 Variation 3: Tapering Idea: the chunk size, K i is a function of not only the remaining work, but also the task cost variance variance is estimated using history information high variance => small chunk size should be used low variance => larger chunks OK See S. Lucco, Adaptive Parallel Programs, PhD Thesis, UCB, CSD , Gives analysis (based on workload distribution) Also gives experimental results -- tapering always works at least as well as GSS, although difference is often small 04/18/2007 CS267 Lecture 24 22

23 Variation 4: Weighted Factoring Idea: similar to self-scheduling, but divide task cost by computational power of requesting node Useful for heterogeneous systems Also useful for shared resource clusters, e.g., built using all the machines in a building as with Tapering, historical information is used to predict future speed speed may depend on the other loads currently on a given processor See Hummel, Schmit, Uma, and Wein, SPAA 96 includes experimental data and analysis 04/18/2007 CS267 Lecture 24 23

24 When is Self-Scheduling a Good Idea? Useful when: A batch (or set) of tasks without dependencies can also be used with dependencies, but most analysis has only been done for task sets without dependencies The cost of each task is unknown Locality is not important Shared memory machine, or at least number of processors is small centralization is OK 04/18/2007 CS267 Lecture 24 24

25 Distributed Task Queues The obvious extension of task queue to distributed memory is: a distributed task queue (or bag ) Doesn t appear as explicit data structure in message-passing Idle processors can pull work, or busy processors push work When are these a good idea? Distributed memory multiprocessors Or, shared memory with significant synchronization overhead Locality is not (very) important Tasks that are: known in advance, e.g., a bag of independent ones dependencies exist, i.e., being computed on the fly The costs of tasks is not known in advance 04/18/2007 CS267 Lecture 24 25

26 Distributed Dynamic Load Balancing Dynamic load balancing algorithms go by other names: Work stealing, work crews, Basic idea, when applied to tree search: busy Each processor performs search on disjoint part of tree When finished, get work from a processor that is still busy Requires asynchronous communication Service pending messages Select a processor and request work idle No work found Do fixed amount of work Got work Service pending messages 04/18/2007 CS267 Lecture 24 26

27 How to Select a Donor Processor Three basic techniques: 1. Asynchronous round robin Each processor k, keeps a variable target k When a processor runs out of work, requests work from target k Set target k = (target k +1) mod procs 2. Global round robin Proc 0 keeps a single variable target When a processor needs work, gets target, requests work from target Proc 0 sets target = (target + 1) mod procs 3. Random polling/stealing When a processor needs work, select a random processor and request work from it Repeat if no work is found 04/18/2007 CS267 Lecture 24 27

28 How to Split Work First parameter is number of tasks to split Related to the self-scheduling variations, but total number of tasks is now unknown Second question is which one(s) Send tasks near the bottom of the stack (oldest) Execute from the top (most recent) May be able to do better with information about task costs Bottom of stack Top of stack 04/18/2007 CS267 Lecture 24 28

29 Theoretical Results (1) Main result: A simple randomized algorithm is optimal with high probability Karp and Zhang [88] show this for a tree of unit cost (equal size) tasks Parent must be done before children Tree unfolds at runtime Task number/priorities not known a priori Children pushed to random processors Show this for independent, equal sized tasks Throw balls into random bins : Θ ( log n / log log n ) in largest bin Throw d times and pick the smallest bin: log log n / log d = Θ (1) [Azar] Extension to parallel throwing [Adler et all 95] Shows p log p tasks leads to good balance 04/18/2007 CS267 Lecture 24 29

30 Theoretical Results (2) Main result: A simple randomized algorithm is optimal with high probability Blumofe and Leiserson [94] show this for a fixed task tree of variable cost tasks their algorithm uses task pulling (stealing) instead of pushing, which is good for locality I.e., when a processor becomes idle, it steals from a random processor also have (loose) bounds on the total memory required Chakrabarti et al [94] show this for a dynamic tree of variable cost tasks works for branch and bound, i.e. tree structure can depend on execution order uses randomized pushing of tasks instead of pulling, so worse locality Open problem: does task pulling provably work well for dynamic trees? 04/18/2007 CS267 Lecture 24 30

31 Distributed Task Queue References Introduction to Parallel Computing by Kumar et al (text) Multipol library (See C.-P. Wen, UCB PhD, 1996.) Part of Multipol ( Try to push tasks with high ratio of cost to compute/cost to push Ex: for matmul, ratio = 2n 3 cost(flop) / 2n 2 cost(send a word) Goldstein, Rogers, Grunwald, and others (independent work) have all shown advantages of integrating into the language framework very lightweight thread creation CILK (Leiserson et al) (supertech.lcs.mit.edu/cilk) Space bound on task stealing X10 from IBM 04/18/2007 CS267 Lecture 24 31

32 Diffusion-Based Load Balancing In the randomized schemes, the machine is treated as fully-connected. Diffusion-based load balancing takes topology into account Locality properties better than prior work Load balancing somewhat slower than randomized Cost of tasks must be known at creation time No dependencies between tasks 04/18/2007 CS267 Lecture 24 32

33 Diffusion-based load balancing The machine is modeled as a graph At each step, we compute the weight of task remaining on each processor This is simply the number if they are unit cost tasks Each processor compares its weight with its neighbors and performs some averaging Analysis using Markov chains See Ghosh et al, SPAA96 for a second order diffusive load balancing algorithm takes into account amount of work sent last time avoids some oscillation of first order schemes Note: locality is still not a major concern, although balancing with neighbors may be better than random 04/18/2007 CS267 Lecture 24 33

34 Mixed Parallelism As another variation, consider a problem with 2 levels of parallelism course-grained task parallelism good when many tasks, bad if few fine-grained data parallelism good when much parallelism within a task, bad if little Appears in: Adaptive mesh refinement Discrete event simulation, e.g., circuit simulation Database query processing Sparse matrix direct solvers 04/18/2007 CS267 Lecture 24 34

35 Mixed Parallelism Strategies 04/18/2007 CS267 Lecture 24 35

36 Which Strategy to Use And easier to implement 04/18/2007 CS267 Lecture 24 36

37 Switch Parallelism: A Special Case 04/18/2007 CS267 Lecture 24 37

38 List of CS267 Projects 04/18/2007 CS267 Lecture 24 38

39 CS267 Projects Matt: Sorting algorithms (UPC impl. & LoPC model) Nutt & Shoiab: parallel liquid simulation for graphic (PIC and parallel sparse poisson solver) Jim: Phylogentic tree construction (neighbor joining algorithm); some aspects run in parallel, but large trees don t fit on single computer (MPI with C) David, Peter, Steve: Parallel driver for iterative refinement for ScaLAPACK Jack: Characterizing and scaling material codes for optimal properties (diag large matrices and calc entries) (MPI with Fortran) JRL: TBD Narayanan & Statish: Statistical model on GPU (NVIDIA) 04/18/2007 CS267 Lecture 24 39

40 More Projects Kyle: DNS simulations of lid-driven cavity flow using Chombo (AMR NavStokes Poisson solvers; change current ring vortex changed to lid; wind on a lake ) (C++ and Fortran with MPI; could you C++ expertise) Max: Adam s unstructured multigrid applied to pumice deformation; may modify to take breakage into account. (C with MPI); maybe room for other partners 04/18/2007 CS267 Lecture 24 40

41 Extra Slides 04/18/2007 CS267 Lecture 24 41

42 Simple Performance Model for Data Parallelism 04/18/2007 CS267 Lecture 24 42

43 04/18/2007 CS267 Lecture 24 43

44 Modeling Performance To predict performance, make assumptions about task tree complete tree with branching factor d>= 2 d child tasks of parent of size N are all of size N/c, c>1 work to do task of size N is O(N a ), a>= 1 Example: Sign function based eigenvalue routine d=2, c=2 (on average), a=3 Combine these assumptions with model of data parallelism 04/18/2007 CS267 Lecture 24 44

45 Actual Speed of Sign Function Eigensolver Starred lines are optimal mixed parallelism Solid lines are data parallelism Dashed lines are switched parallelism Intel Paragon, built on ScaLAPACK Switched parallelism worthwhile! 04/18/2007 CS267 Lecture 24 45

46 Values of Sigma (Problem Size for Half Peak) 04/18/2007 CS267 Lecture 24 46

47 Small Example The 0/1 integer-linear-programming problem E.g., Given integer matrices/vectors as follows: an nxm matrix A, an m-element vector b, and an n-element vector c Find n-element vector x whose elements are 0 or 1 Satisfies the constraint: A*x >= b The function: f(x) = c.dot x should be minimized A = b = 2 c = x 1 + 2x 2 + x 3 + 2x 4 >= 8 (and 2 others inequalities) Minimize: 2x 1 + x 2 x 3 2x Note: 4 24 possible values for x 04/18/2007 CS267 Lecture 24 47

48 Discrete Optimizations Problems in General A discrete optimization problem (S, f) S is a set of feasible solutions that satisfy given constraints. S is finite or countably infinite. f is the cost function that maps each element of S onto the set of real numbers R. The objective of a discrete optimization problem (DOP) is to find a feasible solution x opt, such that f(x opt ) <= f(x) for all x in S. Discrete optimizations problems are NP-complete, so only exponential solutions are known Parallelism gives only a constant speedup Need to focus on average case behavior 04/18/2007 CS267 Lecture 24 48

49 Best-First Search Rather than searching to the bottom, keep set of current states in the space Pick the best one (by some heuristic) for the next step Use lower bound l(x) as heuristic l(x) = g(x) + h(x) g(x) is the cost of reaching the current state h(x) is a heuristic for the cost of reaching the goal Choose h(x) to be a lower bound on actual cost E.g., h(x) might be sum of number of moves for each piece in game problem to reach a solution (ignoring other pieces) 04/18/2007 CS267 Lecture 24 49

50 Branch and Bound Search Revisited The load balancing algorithms as described were for full depth-first search For most real problems, the search is bounded Current bound (e.g., best solution so far) logically shared For large-scale machines, may be replicated All processors need not always agree on bounds Big savings in practice Trade-off between Work spent updating bound Time wasted search unnecessary part of the space 04/18/2007 CS267 Lecture 24 50

51 Simulated Efficiency of Eigensolver Starred lines are optimal mixed parallelism Solid lines are data parallelism Dashed lines are switched parallelism 04/18/2007 CS267 Lecture 24 51

52 Simulated efficiency of Sparse Cholesky Starred lines are optimal mixed parallelism Solid lines are data parallelism Dashed lines are switched parallelism 04/18/2007 CS267 Lecture 24 52